Skip to content

Validate

The validate command checks evaluation files for schema errors without running them.

Terminal window
agentv validate evals/my-eval.yaml
agentv validate evals/greeting.eval.ts

Validate multiple files:

Terminal window
agentv validate "evals/**/*.yaml" "evals/**/*.ts"
  • YAML/JSONL syntax and TypeScript config default exports
  • Required fields: id, input data (input or prompt matrix vars), and at least one grader under assert
  • Grader references (command paths, prompt files)
  • Target references match entries in providers.yaml
  • Rubric structure and field types

For TypeScript eval configs, agentv validate imports and materializes the file through the same core loader used by agentv eval. Supported TypeScript config filenames are *.eval.ts and *.eval.mts; arbitrary .ts helper files are ignored during directory and glob expansion.

  • Before running evaluations to catch config errors early
  • In CI/CD pipelines as a pre-check
  • After editing eval files to verify correctness

agentv validate replaces the old eval mock dry-run use case for schema and configuration checks. It does not execute targets and does not produce quality scores. When you need no-live-LLM quality validation, run against an oracle/reference target or use frozen transcript/replay fixtures so graders see real candidate output.